Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Maruf A. Dhali1 , Sheng He1 , Mladen Popović2 , Eibert Tigchelaar3 and Lambert Schomaker1
1 Institute of Artificial Intelligence and Cognitive Engineering (ALICE), Faculty of Mathematics and Natural Sciences,
University of Groningen, PO Box 407, 9700 AK, Groningen, The Netherlands
2 Qumran Institute, Faculty of Theology and Religious Studies,
Keywords: Dead Sea Scrolls, Handwritten Document Analysis, Digital Palaeography, Writer Identification, Handwriting
Recognition, Pattern Recognition, Feature Representation, Machine Learning.
Abstract: To understand the historical context of an ancient manuscript, scholars rely on the prior knowledge of writer
and date of that document. In this paper, we study the Dead Sea Scrolls, a collection of ancient manuscripts
with immense historical, religious, and linguistic significance, which was discovered in the mid-20th century
near the Dead Sea. Most of the manuscripts of this collection have become digitally available only recently
and techniques from the pattern recognition field can be applied to revise existing hypotheses on the writers
and dates of these scrolls. This paper presents our ongoing work which aims to introduce digital palaeography
to the field and generate fresh empirical data by means of pattern recognition and artificial intelligence. Chal-
lenges in analyzing the Dead Sea Scrolls are highlighted by a pilot experiment identifying the writers using
several dedicated features. Finally, we discuss whether to use specifically-designed shape features for writer
identification or to use the Deep Learning methods on a relatively limited ancient manuscript collection which
is degraded over the course of time and is not labeled, as in the case of the Dead Sea Scrolls.
693
Dhali, M., He, S., Popović, M., Tigchelaar, E. and Schomaker, L.
A Digital Palaeographic Approach towards Writer Identification in the Dead Sea Scrolls.
DOI: 10.5220/0006249706930702
In Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2017), pages 693-702
ISBN: 978-989-758-222-6
Copyright
c 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
1.1 Dead Sea Scrolls image processing techniques need to be applied for
optimum results of segmentation. Starting with edge
The DSS are a collection of ancient damaged detection, morphological operations, filling gaps and
manuscripts that were discovered in the mid-20th cen- then finding connected components help to automati-
tury in the Judaean Desert, in between Jerusalem and cally segment the hand-written fragments. Then fur-
the Dead Sea. Most were written over a period of ther processing can localize and extract the charac-
almost four centuries (ca. 250 BC to ca. 135 AD) ters. Due to the difference in the textures of papyrus
(Tigchelaar, 2010; Popović, 2012; Popović, 2015) in and animal-skin, individual measures must be taken
characters commonly known as the Hebrew alpha- on their distinctive periodic structures.
bet, which actually derives from the Aramaic script We have explored different feature-extraction
(Yardeni, 2002). The manuscripts, written by many techniques on the images of the DSS. Feature-
different writers, some of whom may have written representation maps the raw pixel intensity into a dis-
multiple manuscripts, display a broad variety and de- criminant high-dimensional space (Mikolajczyk and
velopment of different styles of this Hebrew-Aramaic Schmid, 2005; Li et al., 2015) in order to capture
script. The study of ancient handwriting provides specific information of the characters which can be
the chronological framework, but the typological se- processed by algorithms in computers. This step is
quence of writing styles has to date not been system- an important element in the field of computer vision
atically assessed for the DSS. and pattern recognition. There have been a lot of ef-
This project carries out the first systematic assess- forts to design discriminative and powerful features
ment of the palaeographic framework of the scrolls (Li et al., 2015). Though the (Deep) Learning-based
by combining two approaches. First, we will con- feature representation may achieve better results in
duct new radiocarbon (14 C) dating on a number of many cases, hand-crafted features have several advan-
physical samples of the scrolls, kindly provided to tages in the analysis of handwritten documents, espe-
us by the Israel Antiquities Authority (IAA). Sec- cially for historical manuscripts. This is due to the
ond, we will generate for the first time quantitative amount of data in historical manuscript collections,
data for palaeographic handwriting recognition by which is usually not big enough to train deep neural
means of Artificial Intelligence, using the Monk sys- networks. In contrast, the ImageNet data set (Deng
tem, designed by Schomaker’s research group at AL- et al., 2009) contains millions of samples for train-
ICE (Schomaker, 2016; Van der Zant et al., 2008; ing the network. The challenge becomes even higher
Bulacu and Schomaker, 2007). The challenging is- when the total number of usable pages comes to a
sue of writer identification in the DSS has not been count of hundreds in the DSS. To take the opportu-
systematically dealt with before. The tools of digital nities offered by the Deep Learning methods, the as-
palaeography enable new, significant steps forward. sociated challenges need to be overcome in order to
In this paper, we focus on this second approach of analyse the DSS.
digital palaeography.
694
A Digital Palaeographic Approach towards Writer Identification in the Dead Sea Scrolls
posures per fragment. The system then generates a Additionally, the possibility of using the IAA
57th file of a colour image that combines all visible material will improve our results.
wavelengths. The resolution of the files is 1,215 pix-
els per inch at a 1:1 ratio, capturing approximately 2.2 Ground Truth
4 gigabytes of information per fragment (Shor et al.,
2014). Unlike many other historical manuscripts, the DSS
The Brill images are single-layered grayscale im- do not have a structured and complete dataset nor the
ages with 300 ppi (pixels per inch) on both axis. They ground truths for testing. Before diving into any sort
have shadows and reflection from external lighting. of computer aided writer identification, the ground
Additionally the lighting throughout all the images is truths must be there to analyse the results. To estab-
not uniform. Among the images, the ones containing lish the ground truths, we need experts in the field
several fragments are mostly not aligned in a horizon- and also their proper access to the data. We have this
tal way for text reading. This poses the issue of rota- two-folded advantage in our group: first through the
tion variance in characters. Many of the images also presence of palaeographic experts and second through
contain paper calibration strips for scale representa- the Monk system which is accessible through web
tion and contemporary hand-written numbers. The browsers. By integrating these, we started to label the
digitization noise can also be noticed in many of the DSS image data for ground truths.
images (see figure 2). We have proposed two different methods for la-
belling. The first one is to detect the region of interest
in the DSS images. The second one is to create the
ground truth for character-labels. Both these tasks re-
quire manual labour from experts with palaeographic
background knowledge on image level to pixel level.
This is the bench-mark in identifying the writers and
aligning the temporal developments in script style
Figure 2: Two of the Brill images; PAM 40.456 (left) and for the DSS. In this paper, we will only use the la-
PAM 40.531 (right); the images show digitisation noise, belled regions of interest (we call them FragmROIs,
alignment issues with the small fragments, shadows near by shortening the term fragment region of interest)
the border and lighting problems. in order to build algorithms to extract features (using
available methods) and identify (recognition) writers.
The IAA images are clear, properly aligned and From the DSS images, FragmROIs were selected and
free from the problem of lighting and shadow un- labelled by the palaeographic experts using the Monk
like the Brill ones. Additionally the different expo- system (see figure 3). Those rectangular FragmROIs
sure bands of the IAA hold important underlying in- could consist of the entire text on an image, or of only
formation regarding the fragments providing essential a section of text selected from an image. Different
attributes for the scrolls. For example, one particu- FragmROIs from one and the same manuscript were
lar band provides clear information on the ink (fore- labelled as stemming from one writer or scribe, unless
ground) whereas another one gives more details on palaeographers distinguished two scribes as writers of
the underlying leather/papyrus (background) on the the manuscript.
retro side. Some bands are useful for the textual con-
tents and some other bands give a better understand-
ing on the textural properties of the scroll material.
Extraction of this useful information is possible on
both single images and multi-spectral-fused images.
As a whole, digital image data has provided a new
and broader perspective in the quantitative analysis
and processing of the scrolls.
The scope of the current paper is limited to the Figure 3: Using the Monk system, the palaeographic expert
images in the Brill collection, but in the near future can select the region of interest (FragmROI), then put the
we expect to publish our results on the digital data associated attributes (scribe, style, comment etc.) to pro-
from the IAA’s Leon Levy Dead Sea Scrolls Digital duce an XML file, which will later be used as labelled data.
Library. The quality and the challenges of the Brill Currently the experts can only select a rectangular region,
but the provision of choosing a polygonal region of interest
images can not be seen as a set-back, rather it would will be added to the system in near future.
be a starting benchmark to the robustness of our work.
695
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
While labelling the writers, we have set up a provi- this section, we will present the methodology used in
sional naming rule starting with the name scribeAxxx, our pilot project in writer identification to benchmark
where xxx are numerical values starting from 001. our works in analysing the DSS.
Each of the human labelled new-writers will be al-
located with an individual value. The term A is put 3.1 Writer Identification
before the numerical values in order to preserve the
tag of original labelling from the palaeographic ex- Identifying writers using computers has been done for
perts. If at a later stage of our study, two of the decades (Plamondon and Lorette, 1989), which is a
writers are found to be the same one according to problem of recognizing the writer of a given docu-
the system, then they can be referred to with the new ment based on handwriting styles. A number of dif-
name of scribeBxxx having two child-node of format ferent features have been proposed and studied for
scribeAxxx, preserving the original label. writer identification on scripts from several languages
The present pilot study is based on two distinct including Dutch (Bulacu and Schomaker, 2007), En-
set of writers. The first set is a limited sample of glish (Schomaker and Bulacu, 2004), Indic (Adak
323 FragmROIs labelled as having been written by and Chaudhuri, 2015; Karunakara and Mallikarju-
13 scribes, namely the scribes of 1QIsaa columns naswamy, 2011), and Arabic (Bulacu et al., 2007).
1-27, 1QIsaa columns 28-54, 1QS, 1QSa, 1QSb, In the case of the DSS, we will be identifying the
1QM, 1QpHab columns 1-12, 1QpHab columns 12 scribes behind the scrolls with Hebrew characters, and
end-13, 4Q53, 4Q175, 11Q5, 11Q19 columns 2- a hand-crafted feature specially for these characters is
5, 11Q20. We labelled them from scribeA001 to yet to be proposed and studied. Instead of design-
scribeA013. Distinct manuscripts were labelled as ing a new feature, we initially started working with
deriving from different writers, even though in sev- some of the existing textural-based and grapheme-
eral of the manuscripts of the first set palaeographers based features. Textural-based features are based on
think that one and the same writer produced multi- the statistical information about slant and curvature
ple manuscripts (Tigchelaar, 2002). To incorporate of the handwritten characters, and grapheme-based
the palaeographic opinion, we then merged those 13 features, inspired by the bag-of-words model, extract
scribes into 7 scribes by introducing scribeBxxx se- local structures and then map them into a common
ries. Then we took the second set of 13 scribes space (He and Schomaker, 2016). We briefly discuss
with a limited sample of 124 FragmROIs labelled the preprocessing techniques and the features used in
as scribeA014 to scribeA026 (the scribes of 4Q266, this work in the following sections.
4Q504, 11Q10, 1Q22, 4Q209, 4Q167, 4Q6, 4Q286,
4Q381, 4Q405, 4Q491, 4Q431, 4Q525). The main 3.1.1 Preprocessing
difference between these two sets are the amount of
characters per scribe. The first set has a higher num- As the feature extraction technique is applied on the
ber of characters than the second set. Thus, for this binarized images, first we pre-processed the DSS im-
pilot project, we have 447 FragmROIs labelled as ages. Binarizing the Dead Sea Scrolls images is quite
20 distinguishable scribes according to palaeographic challenging, given their diverse intensity, similarity
opinion. between ink and background traces, and image qual-
ity. We first started with Sobel edge detection (So-
bel, 1990) and then removed the connected objects on
the border to get rid of the markings. Morphological
3 A PILOT PROJECT operation was then used followed by image thresh-
olding. We used the global Otsu threshold selection
The DSS image data has its own distinctive char- method (Otsu, 1975) as it is efficient and parameter-
acteristics compared to other historical manuscript less (see figure 4).
datasets. This data set has with quite a different ap-
pearance from, e.g. historical manuscripts such as the 3.1.2 Feature Representation
Medieval Palaeographic Scale data set (Monk, 2016)
from a previous project (He et al., 2016) in three as- Previous studies showed that the textural-based
pects: (1) the number of characters in fragments from feature extraction methods perform better than
some documents can be as low as one; (2) the ink of grapheme-based methods (He and Schomaker, 2016;
each character has been faded out over the course of He and Schomaker, 2017). Additionally, a more pow-
time, making it more difficult to observe and process; erful approach was introduced by using the spatial co-
(3) the large diversity and lack of uniformity among occurrence among features (Bulacu and Schomaker,
text blocks, presenting a challenge for analysis. In 2007; Ito and Kubota, 2010; Qi et al., 2014). The
696
A Digital Palaeographic Approach towards Writer Identification in the Dead Sea Scrolls
697
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
698
A Digital Palaeographic Approach towards Writer Identification in the Dead Sea Scrolls
put FragmROI. The top-n performance is calculated Table 3: The top-1 and top-10 performance (in percent-
when the query FragmROI is recognized as the writer age) of writer identification for another 13 scribes from
of the FragmROI on the top n of the hit-list. For ex- scribeA014 to scribeA026, with limited text fragments.
ample, the top-10 hit-list signifies the overall percent- Feature Top-1 Top-10
age of finding the same writer as input within the first Hinge 61.90 92.06
ten candidates (shortest distanced) of the output hit- CoHinge 62.69 89.68
list. Similarly the top-1 means the top-most candidate ∆1 Hinge 43.65 85.71
in the output hit-list corresponds to the same writer QuadHinge 63.49 90.47
as the input. The performance of top-1 and top-10 Quill 48.38 89.51
hit-list for first set is presented in Table 1. Accord- QuillHinge 45.16 74.19
TripleChainCode 61.90 88.09
Table 1: The top-1 and top-10 performance (in percentage) COLD 58.87 88.71
of writer identification for 13 scribes from scribeA001 to
scribeA013.
Junclet 31.45 76.74
Feature Top-1 Top-10 Table 4: The top-1 and top-10 performance (in percentage)
Hinge 87.61 97.83 of writer identification for all 20 scribes.
CoHinge 81.11 95.97
Feature Top-1 Top-10
∆1 Hinge 79.87 94.73 Hinge 78.30 94.40
QuadHinge 89.47 96.59 CoHinge 79.19 89.93
Quill 80.80 93.80
∆1 Hinge 68.23 85.48
QuillHinge 76.78 89.78
QuadHinge 79.64 89.04
TripleChainCode 84.82 96.59
Quill 71.58 86.57
COLD 82.35 94.42
QuillHinge 69.57 82.10
Junclet 81.42 95.04
TripleChainCode 79.19 91.28
ing to the majority palaeographic opinion, scribeA001 COLD 76.95 88.37
and scribeA002 are the same scribe, and so also Junclet 40.71 72.93
scribeA003, 004, 005, 008, 009 and scribeA010, 012,
which are then labelled as scribeB001, B002 and
B003 respectively. The result is presented in Table 2
for these seven scribes. Then we took the second set
5 DISCUSSIONS
Table 2: The top-1 and top-10 performance (in percentage) 5.1 Performance Evaluation
of writer identification for 7 scribes: scribeB001, B002,
A006, A007, B003, A011 and A012 .
We presented the results of writer identification on a
Feature Top-1 Top-10 limited number of scribes. Of the shape based meth-
Hinge 92.26 98.76 ods, the QuadHinge performs the best in the top-1
CoHinge 93.80 97.52 hit-lists for three out of four cases (only for the case
∆1 Hinge 90.71 96.28 of seven scribes in Table 2, CoHinge performs better
QuadHinge 93.50 96.90 with a small difference of 0.30% than QuadHinge),
Quill 88.54 96.28 whereas the Hinge feature gives better result in all
QuillHinge 88.85 96.90 the top-10 hit-lists. The reason for this performance
TripleChainCode 92.26 98.14 can be deduced from the design criteria of the fea-
COLD 91.33 96.28 tures themselves. Hinge feature takes into account
Junclet 56.03 88.85 the joint probability distribution of the orientations of
legs of two contour fragments from a common end
of 13 scribes with 124 FragmROIs. The amount of pixel on ink contours, which proves to be a strong
text is lower in this set than in the first one. The identical property for individual scribes of these an-
result is shown in Table 3. Finally, we took all the cient manuscripts. Additionally, the incorporation of
scribes together for testing. Table 4 presents the result FCM to the Hinge feature following the JFD principle
of these 20 scribes together ( i.e., scribeB001, B002, gives the QuadHinge feature a boosted performance.
B003, A006, A007, A011, A013, A014 to A026). We The directional measurement of the ink-trace
briefly discuss the results and our propositions in the width makes the Quill feature, which is quite infor-
next section (Section 5). mative on quill-based medieval scripts, a weak can-
699
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
didate for the DSS. This is due to the uniformity of networks on massive not-labelled handwriting collec-
the ink-trace in these documents coming from a prob- tions. Such networks are trained to reconstruct im-
ably fairly blunt tip of the ancient writing equipment. ages over (via) a very limited number of values (hid-
Consequently, the QuillHinge fails to provide a higher den units). After training, such a network implicitly
performance in this test set. The ∆1 Hinge has a lim- knows a lot about historical handwritings in general.
ited performance, indicating that on Hebrew charac- In a second stage, such a network is then applied to
ters, loss of the angle with respect to the horizontal the DSS, using those hidden unit vectors as feature
removes too much of the writer-specific information. descriptors.
The grapheme-based feature, Junclets, gives lower 3) Data augmentation can be utilized in the pro-
performance than the cross-script writer identification cessing of the DSS. If there is a believable random
(He et al., 2015) due to the lower variability in the transformation of the DSS’ text patterns, i.e., one that
stroke-length distribution in every direction around a remains legible by humans, then for each natural sam-
reference point inside the ink of the DSS’ Hebrew ple of a character, a number of N derived random ver-
characters. sions of it may be added to the training set, effectively
enlarging the amount of labelled data. Known already
5.2 Propositions in the nineties (Baird, 1992) this was later made pop-
ular in handwriting recognition later by the use of hid-
The challenges in analysing the DSS are unique den Markov models (Varga and Bunke, 2003; Ha and
and unprecedented. Using the dedicated features Bunke, 1997).
(in 3.1.2), we found fast results without lengthy train-
ing on the limited labelled data of the DSS. But 5.3 Conclusions
they are certainly not the best results to be expected.
Especially when the amount of data is small with In this paper, we have introduced digital palaeogra-
large variability (as in Table 3), the performance be- phy of the DSS by presenting a pilot project, which
comes lower. To overcome this situation, we need to is part of a pioneering multi-disciplinary project that
consider a pragmatic approach incorporating several brings together the natural sciences, artificial intel-
propositions. ligence, and the humanities. By introducing the
1) Statistical modelling can be used in the case rule to establish ground-truths, we performed writer-
of the DSS, where the sample size is low and there identification tests using dedicated features on provi-
are differences in the scholarly opinion of writers as sionally labelled data. The varying performance of
well. We can use the differences in writing attributes results for different sets of writers led us to the propo-
of a set of different manuscripts to build a population sitions of statistical modelling, transfer learning, and
model. A writer model can be built using the query data augmentation for this largely diverse collection
manuscript. The classification is then carried out by of manuscripts.
evaluating the similarity of a further manuscript sam- We consider the results of this paper as a baseline
ple with respect to the models. We can build our pro- measurement for our later experiments. We will com-
visional model, similar to the work of speaker identi- bine both the aspect of specifically-designed shape
fication (Leuzzi et al., 2016), as follows: features and the Deep Learning methods to produce
pb (d(Wi ,W j )) fresh empirical data for the study of the DSS. Addi-
Λ(d(Wi ,W j )) = (4) tionally, we will conduct new radiocarbon (14 C) dat-
pw (d(Wi ,W j ))
ing on a number of physical samples of the scrolls.
Here, d(Wi ,W j ) is the distance computed from Wi , The outcome of 14 C dating will then be subjected to
the query writer to W j , the suspected writer. Λ de- Bayesian statistics methods in combination with the
notes the likelihood ratio over d(Wi ,W j ). The distri- results from temporal alignment using pattern recog-
bution of distances between the suspected writer and nition to reach more accurate and precise dating of the
the population is denoted by pb (d(Wi ,W j )), which DSS.
can be referred as the between-group distance among
the writers. pw (d(Wi ,W j )) is the distribution of dis-
tances taken within different instances of the sus-
pected writer (within-group distance). The collection ACKNOWLEDGEMENTS
of statistical models (Fisher, 1925), analysis of vari-
ance (ANOVA), can be used to analyse the within- The authors would like to thank Ruwan van der Iest
group and between-group variances of the writers. (research assistant for the ERC project at the Qum-
2) Another possibility is transfer learning (Long ran Institute) for his valuable inputs in labelling the
and Wang, 2015). It starts with the use of pre-trained regions of interest through the Monk-system.
700
A Digital Palaeographic Approach towards Writer Identification in the Dead Sea Scrolls
This work has been supported by an ERC Starting Ito, S. and Kubota, S. (2010). Object classification us-
Grant of the European Research Council (EU Hori- ing heterogeneous co-occurrence features. In Euro-
zon 2020): The Hands that Wrote the Bible: Digital pean Conference on Computer Vision, pages 701–714.
Palaeography and Scribal Culture of the DSS (Hand- Springer.
sandBible # 640497). Additional support comes from Karunakara, K. and Mallikarjunaswamy, B. (2011). Writer
identification based on offline handwritten document
NWO (Netherlands Organisation for Scientific Re- images in kannada language using empirical mode de-
search) and FWO (the Research Foundation Flan- composition method. Writer, 30(6).
ders): Models of Textual Communities and Digital LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learn-
Palaeography of the DSS (# 326-25-001). ing. Nature, 521(7553):436–444.
Leuzzi, F., Tessitore, G., Delfino, S., Fusco, C., Gneo, M.,
and Zambonini, G. (2016). A statistical approach to
REFERENCES speaker identification in forensic phonetics field.
Li, Y., Wang, S., Tian, Q., and Ding, X. (2015). Feature
Adak, C. and Chaudhuri, B. B. (2015). Writer identification representation for statistical-learning-based object de-
from offline isolated bangla characters and numerals. tection: A review. PR, 48(11):3542–3559.
In ICDAR, pages 486–490. IEEE. Lim, T. and Alexander, P. (1995). Volume 1. In The Dead
Baird, H. S. (1992). Document image defect models. In Sea Scrolls Electronic Library. Brill.
Structured Document Image Analysis, pages 546–556. Long, M. and Wang, J. (2015). Learning transferable
Springer. features with deep adaptation networks. CoRR,
Belongie, S., Malik, J., and Puzicha, J. (2002). Shape abs/1502.02791, 1:2.
matching and object recognition using shape contexts. Mikolajczyk, K. and Schmid, C. (2005). A perfor-
IEEE PAMI, 24(4):509–522. mance evaluation of local descriptors. IEEE PAMI,
Benhamou, S. (2004). How to reliably estimate the tortu- 27(10):1615–1630.
osity of an animal’s path: straightness, sinuosity, or Monk (2016). Medieval palaeographic scale data set (online
fractal dimension? Theoretical Biology, 229(2). collection).
Brink, A., Smit, J., Bulacu, M., and Schomaker, L. (2012). Otsu, N. (1975). A threshold selection method from gray-
Writer identification using directional ink-trace width level histograms. Automatica, 11(285-296):23–27.
measurements. PR, 45(1):162–171. Plamondon, R. and Lorette, G. (1989). Automatic signature
Bulacu, M. and Schomaker, L. (2007). Text-independent verification and writer identification - the state of the
writer identification and verification using textural and art. Pattern recognition, 22(2):107–131.
allographic features. IEEE PAMI, 29(4):701–717. Popović, M. (2012). Qumran as scroll storehouse in times
Bulacu, M., Schomaker, L., and Brink, A. (2007). Text- of crisis? a comparative perspective on judaean desert
independent writer identification and verification on manuscript collections 1. Journal for the Study of Ju-
offline arabic handwriting. In ICDAR, volume 2, daism, 43(4-5):551–594.
pages 769–773. IEEE. Popović, M. (2015). The ancient ‘library’ of qumran be-
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei- tween urban and rural culture. In The Dead Sea Scrolls
Fei, L. (2009). ImageNet: A Large-Scale Hierarchical at Qumran and the Concept of a Library, pages 155–
Image Database. In CVPR. 167. Brill.
Fisher, R. A. (1925). Statistical methods for research work- Prasad, D. K., Quek, C., Leung, M. K., and Cho, S.-Y.
ers. Genesis Publishing Pvt Ltd. (2011). A parameter independent line fitting method.
Ha, T. M. and Bunke, H. (1997). Off-line, handwritten In ACPR, pages 441–445.
numeral recognition by perturbation method. IEEE Qi, X., Xiao, R., Li, C.-G., Qiao, Y., Guo, J., and Tang,
PAMI, 19(5):535–539. X. (2014). Pairwise rotation invariant co-occurrence
He, S., Samara, P., Burgers, J., and Schomaker, L. (2016). local binary pattern. IEEE PAMI, 36(11):2199–2213.
A multiple-label guided clustering algorithm for his- Schomaker, L. (2016). Design considerations for a large-
torical document dating and localization. IEEE Trans- scale image-based text search engine in historical
actions on Image Processing, 25(11):5252–5265. manuscript collections. it-Information Technology,
He, S. and Schomaker, L. (2014). Delta-n hinge: Rotation- 58(2):80–88.
invariant features for writer identification. In ICPR, Schomaker, L. and Bulacu, M. (2004). Automatic writer
pages 2023–2028. identification using connected-component contours
He, S. and Schomaker, L. (2016). Writer identification us- and edge-based features of uppercase western script.
ing curvature-free features. PR. IEEE PAMI, 26(6):787–798.
He, S. and Schomaker, L. (2017). Beyond ocr: Multi- Shor, P., Manfredi, M., Bearman, G. H., Marengo, E.,
faceted understanding of handwritten document char- Boydston, K., and Christens-Barry, W. A. (2014). The
acteristics. PR, 63:321–333. leon levy dead sea scrolls digital library: The digitiza-
He, S., Wiering, M., and Schomaker, L. (2015). Junction tion project of the dead sea scrolls. Journal of East-
detection in handwritten documents and its applica- ern Mediterranean Archaeology and Heritage Studies,
tion to writer identification. PR, 48(12):4036–4048. 2(2):71–89.
701
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
702